In Proceedings of the Chicago Linguistics Society (CLS 38), ed. by M. Andronis,
E. Debenport, A. Pycha and K. Yoshimura, pp. 79-90, 2002.
Seem and other hazards of the lexicon
Misha Becker
University of Pennsylvania and UNC Chapel Hill
1 Introduction
This paper is about how children learn the class of verbs known as raising verbs.
This class includes verbs such as seem, appear, tend (to), happen (to), used (to),
the adjectival predicate be likely and a few others. Raising predicates are defined
as a class by their failure to select any NP arguments. They are like modals or
auxiliary verbs in this respect, but raising verbs in English are unlike modals and
auxiliaries in their morphosyntax (raising verbs inflect like main verbs do). In
languages that require main clauses to have an overt subject, such as English, the
failure of these predicates to select a subject argument means that the NP subject
of the lower clause must raise to the subject position of the main clause (or else
an expletive subject must be inserted).
(1)
[IP e [I’ [VP [V seems [IP John [to be happy]]]]]]
Raising predicates are interesting from a developmental perspective
because they present certain difficulties, or “hazards” for the language learner.
Moreover, their acquisition has not yet been studied systematically. To see how
these verbs are hazardous, let us first review the learning of normal, nonhazardous verbs.
1.1 Learning verbs
A considerable body of work on children’s verb learning has shown that in
learning the meanings of verbs, children exploit evidence from various
information sources (please see Gleitman, 1990 for discussion). One source is
observation of the world (i.e. what is going on in the world when a particular verb
is uttered?). Another source is semantic cues from the sentence. For instance,
knowing the meanings of the NPs that occur with a verb can provide possible cues
to a verb’s meaning, or at least exclude highly unlikely meanings. For example,
given the meanings of the subject and object NPs in (3), the novel verb could
mean ‘eat’ or ‘bake’ but is unlikely to mean ‘kiss’.
(2)
(3)
World cues: observe a hitting event + hear the verb hit uttered
Semantic cues: the chef gorped the cake → gorp = eat, bake, #kiss
Another type of information source children make use of is syntactic
information, that is, information from the subcategorization frames in which a
verb can occur. For instance, a verb that occurs in a ditransitive frame (as in (4a))
might mean something about transfer, such as “put” or “give”, but it cannot mean
something like “sleep”, “hit” or “think”. A verb that occurs in an intransitive
frame (4b) might mean something like “sleep” but could not mean “hit” or “put”.
And a verb that occurs with a sentential complement (4c) might mean something
about a mental state, such as “think”, but could not mean “hit”, “put” or “sleep”.
This is so because there are regularities in the mapping between the lexical
meaning of verbs and verbs’ argument structure or subcategorization frames
(Chomsky 1981, Jackendoff 1983, Fisher, Gleitman & Gleitman 1991).
(4)
Syntactic cues:
a. NP V NP PP → put, give, *sleep, *hit, *think
b. NP V → sleep, *hit, *put
c. NP V CP → think, know, *sleep, *hit, *put
This syntactic information is known to be used by both children and adults
in learning the meanings of novel verbs (Fisher et al. 1994, i.a.). Although
information about a verb’s syntactic frame(s) does not point to the precise
meaning of a verb, it is extremely useful in that it excludes a range of impossible
meanings, thereby narrowing the hypothesis space for the learner.
While none of these information sources taken alone can provide the
learner with all the necessary information for learning a verb’s meaning, taken
together they can provide strong cues for the learner. Moreover, these different
kinds of information sources are used to different degrees for different kinds of
verbs: with more concrete verbs (hit) observation of the world might be a very
useful source, while for more abstract verbs (think) it is not. Abstract verbs,
however, are very well cued by syntactic frame information (there is a small range
of meanings associated with the verbs that take sentential complements, e.g.
think), while more concrete verbs tend to be less well cued by syntax (there is a
wide variety of meanings compatible with the transitive frame) (Kako 1998).
1.2 Learning raising verbs
Now let us return to our original question, which is how raising verbs are learned.
In light of what we know about learning normal verbs, what is interesting about
raising verbs is that their information sources are impoverished on all three fronts.
Observational cues are very poor: if you ask what is going on in the world when
seem is uttered, it could be anything; and it is hard, if not impossible, to observe
seeming.
Given this, and given Kako’s observation that abstract verbs are well cued
by syntax, we would hope that there are very good syntactic and semantic cues to
the meanings of raising verbs. But again raising verbs present problems: as
mentioned above, they do not select their subject, so they don’t stand in a
semantic relationship with their syntatic subject (the NP subject of the main
clause). This is because raising verbs don’t assign any theta-roles. Thus, knowing
what the subject means will not point to or exclude anything about the meaning of
the raising verb. (In other words, John and seems do not stand in a semantic
relationship in (1) or in (5a).)
Moreover, the syntax of raising constructions is not transparent. Raising
verbs overlap with control verbs in one kind of sentence frame, namely that in (5).
(5)
a. John seems to be happy.
b. John wants to be happy.
These sentences are string-identical but are generated by very different structures,
illustrated in (6).
(6)
a. [IP [DP Johni] [I’ [VP [V seems] [IP ti to be happy] ] ]
b. [IP [DP Johni] [I’ [VP [V wants] [IP PROi to be happy] ] ]
Unlike in the raising sentence in (5a/6a), the subject of the control
sentence in (5b/6b) is an argument of the main clause verb, so it stands in a
semantic relationship with the verb (wants). The problem for the learner, then, is
that of determining, on the basis of string input, what the correct structure is for a
given string. To put this in perspective, if you hear the sentence
(7)
John gorps to be happy,
how do you know which of the structures in (8) is the correct structure for (7)?
(8)
a. [IP [DP Johni] [I’ [VP [V gorps] [IP ti to be happy] ] ]
b. [IP [DP Johni] [I’ [VP [V gorps] [IP PROi to be happy] ] ]
Thus, we can think about this learning problem as the problem of how a
learner maps a string of words onto the correct structure, and also as the problem
of how a learner determines the nature of silent categories (is a null subject trace
or PRO?). There may be learnability theoretic reasons why a learner should
hypothesize one structure over the other (Borer & Wexler 1987, Frank 1998). But
even if the learner does have a bias to parse (7) as either (8a) or (8b), the learner
still must entertain both structures and must determine which lexical items occur
in structure (8a) and which in (8b). Thus, the question of what information is
available in the input that could lead a learner to solve this mapping problem (and
hence to converge on the adult grammar) remains an interesting question. It is this
question I am addressing in this paper.
2 Experiments with adults
Here I report on the results of a series of experiments I have conducted with adult
speakers of English, aimed at finding out what cues are available in sentences that
might suggest that a verb is a raising verb. The general method I have employed is
to ask adult speakers to fill in a blank in a sentence with a word that would make
the sentence sound natural. I take subjects’ responses to be indicative of the
structure they assign to the sentence. Thus, if a subject responds with a verb like
seem in the sentence John _____ to be happy, then I assume the person assigned
to that sentence a raising structure. If someone responds with want in the same
sentence, then I assume they assigned to the sentence a control structure.
In each experiment, each subject was given 4 test items plus about 40 filler
sentences.i Filler sentences called for nouns, adjectives, transitive and intransitive
verbs, and in some experiments modals and adverbs. Subjects were provided with
the part of speech of the missing word. All subjects were native English speakers
(they were undergraduates at the University of Pennsylvania) and received course
credit for their participation. There were 20 subjects in each experiment.
2.1 Experiment 1: Frames I
Raising and control verbs overlap in the sentence frame John ____ to be happy,
but they can be disambiguated. Raising verbs can occur with an expletive subject
(since they don’t select a subject argument), but control verbs cannot.
(9)
a. It seems that John is happy.
b. *It wants that John is happy.
We might then predict that given a sentence with this frame, subjects will
analyze the missing verb as a raising verb. Some examples of sentences in
Experiment 1 are given in (10).
(10)
a. It _______ that Sophie was upset.
b. Sophie ______ to be upset.
(unraised frame)
(raised frame)
Results of Experiment 1 are given in Table 1.
Table 1. Subjects’ responses in Experiment 1
response type
unraised frame
raised frame
Raising verb
42.5%
35%
Control verb
0%
57.5%
Other
57.5%
7.5%
No control verbs were offered in the unraised frame. This is unsurprising
since control verbs are ungrammatical in this context. What is suprising given our
predictions, however, is that raising verbs were not the most frequent response in
this frame. Moreover, raising verbs were not offered significantly more in the
unraised frame than in the raised frame. Thus, the unraised frame (It ___ that…)
does not unambiguously point to a raising verb. Instead, subjects offered two
kinds of “other” responses. One kind was verbs like say, know or feel; these are
verbs which select a thematic subject argument and take a sentential complement,
as in (11).
(11)
It said that Robert loved Zoe.
Thus, subjects who offered this kind of response were analyzing the it
subject as a referring pronoun, not as an expletive.
The other kind of response was verbs like show, help or reveal, as in (12).
(12)
It showed that Benjamin drove his mother crazy.
In this case, the it subject is either an expletive or a pronoun with very
abstract reference. But in any case, these verbs are clearly not raising verbs. For
example, (13) is ungrammatical.
(13)
*Benjamin showed [t to drive his mother crazy]
Since an it subject is not unambiguously an expletive in the test items in
Experiment 1, in Experiment 2 I compared sentence frames with expletive and
referential it.
2.2 Experiment 2: Frames II
Examples of sentences in this experiment are given in (14).
(14)
a. It ______ to be raining.
b. It ______ to fit in the hole.
(expletive)
(referential)
Again, subjects were asked to fill in the blank. Results are given in Table 2.
Table 2: Subjects’ responses in Experiment 2
response type
expletive it
referential it
Raising verb
83.3%
31.7%
Control verb
3.3%
21.7%
Other
13.3%
46.7%
When an it subject was an unambiguous expletive (i.e. it occurred with a
weather predicate), subjects offered a raising verb response 83.3% of the time,
significantly more than any other type of response. But when the it subject was
referential, subjects offered raising and control responses at roughly equivalent
rates. (Subjects’ “other” responses with an expletive subject were ungrammatical,
e.g. It sucks to be raining; “other” responses with referential it were largely
purpose constructions, e.g. It folded to fit in the hole.) Thus, the combination of an
it subject and a weather predicate serves as a good cue that the main clause verb is
a raising verb.
In a further experiment (Experiment 2a) I gave subjects sentences
containing expletive there. Examples of the test sentences in this experiment are
given in (15).
(15)
a. It ______ to be raining for most of the morning.
b. There ______ to be no end to his complaints about the situation.
The results are shown in Table 3.
Table 3: Subjects’ responses in Experiment 2a
response type
expletive it
expletive there
Raising verb
90%
97.5%
Control verb
0%
0%
Other
10%
2.5%
Thus, expletive there appears to be an even stronger cue to a raising verb
than expletive it (though the difference in Experiment 2a between raising
responses to it and to there is not significant).
2.3 Experiment 3: Eventivity
Experiment 2 showed that adults can identify raising verbs on the basis of certain
sentence frame information, in particular occurrence with an expletive subject.
But returning to our original problem, the learner has to determine the structure of
the ambiguous frame (John gorped to be happy) on the basis of string input.
While there is evidence that children reason across frames (Naigles, Gleitman &
Gleitman 1989, Naigles 1996), and might be able to use information about a
verb’s occurrence with an expletive to make a guess about that verb’s syntactic
properties upon hearing it in another frame, it would still be nice if there were
information from the single ambiguous frame to suggest to a learner that the
structure is a raising or a control structure. Is there such information?
I claimed above that raising verbs do not select any arguments, and this is
true. But they do select something: they select a propositional complement. And
they appear to have preferences about certain aspects of that complement. In
particular, raising verbs appear to prefer a stative predicate inside their
complement.
(16)
a. ??John seems to eat an apple (right now).
b. John seems to be eating an apple (right now).
c. John seems to know the answer (right now).
(eventive)
(stative)
(stative)
In this experiment, I gave subjects the ambiguous sentence frame but
manipulated the eventivity of the lower predicate. Examples are given in (17).
(17)
a. Peter ______ to throw lima beans.
b. Peter ______ to be throwing lima beans.
Results are given in Table 4.
(eventive)
(stative)
Table 4: Subjects’ responses in Experiment 3
response type
eventive predicate
Raising verb
2.5%*
Control verb
60%
Other
37.5%
*p ≤ 0.001
stative predicate
40%*
52.5%
7.5%
In fact, subjects gave virtually no raising verbs as responses when the
lower predicate was eventive (to throw), but they gave a raising verb 40% of the
time when it was stative (to be throwing). In addition to progressive verbs, I ran
the same experiment but with bare stative verbs (e.g. know, be tall), and the types
of responses to bare stative verbs were much like the responses to sentences with
progressive verbs: on average, subjects offered a raising verb 30% of the time, a
control verb 60% of the time, and something else 10% of the time.ii
While subjects offered a raising verb far more with a stative lower
predicate than with an eventive lower predicate, they still offered more control
verbs than raising verbs when the sentence contained a stative lower predicate.
But in all of the test items in this experiment the subject of the sentence was
animate. Animate things have volition and intention, so they can want or try to do
things. Inanimate things cannot. In the final experiment, I manipulated both the
animacy of the subject and the eventivity of the lower predicate.
2.4 Experiment 4: Animacy and eventivity
Examples of test items in this experiment are given in (18-19).
(18)
(19)
Animate subject
a. The driver ______ to hit the car.
(eventive)
b. His campaign manager _____ to remain a problem for the mayoral
candidate.
(stative)
Inanimate subject
a. The boulder ______ to hit the car.
(eventive)
b. The extramarital affair _____ to remain a problem for the mayoral
candidate.
(stative)
Looking first only at animacy, and collapsing across the predicate types, we find
that animate subjects do favor a control verb response, while inanimate subjects
appear to favor a raising verb response. This result is shown in Table 5:
Table 5: Subjects’ responses in Experiment 4: Animacy only
response type
animate subject
inanimate subject
Raising verb
18.75%*
43.75%*
Control verb
51.25%*
17.5%*
Other
30%
38.75%
*p ≤ 0.01
But looking more closely and teasing apart both animacy and eventivity,
we see an interesting interaction. As in Experiment 3, when the subject is animate
we find almost no raising verbs in sentences with an eventive predicate, and some
raising verbs in sentences with a stative predicate. But when the subject is
inanimate we see an interesting pattern: when the predicate is eventive, we mostly
get “other” kinds of responses (mostly purpose constructions, e.g. The wind blew
to lift the leaves). When the predicate is stative, we get mostly raising verbs: in
fact we get almost as many raising verb responses in this condition as we did with
an expletive it subject and a weather predicate in Experiment 2. The proportions
are given in Table 6.
Table 6: Subjects’ responses in Experiment 4: Animacy and eventivity
response type
animate subject
inanimate subject
eventive
stative
eventive
stative
Raising verb
5%*
32.5%*
17.5%*
70%*
Control verb
62.5%
40%
32.5%
2.5%
Other
32.5%
27.5%
50%
27.5%
*Significant interaction for raising verb response: p = 0.038
Thus, the pairing of an inanimate main clause subject with a stative downstairs
predicate is a very good cue that the sentence has a raising structure.
2.5 Summary of results
The five experiments described above were designed to find out, by looking at
adult English-speakers’ responses to the fill-in-the-blank task, what information is
available in the input to cue a learner to a raising verb (as opposed to a control
verb or another kind of verb). We saw that although raising and control verbs are
disambiguated by the frame It ___ that …, that frame does not provide an
unambiguous cue that the main verb is a raising verb. That is, an it subject alone
is not enough, because speakers often interpret the it subject as a referring
pronoun. But forcing it to be an expletive or using expletive there yields a much
higher raising verb response.
We also saw that even though raising verbs do not select any arguments
and therefore do not stand in a semantic relation with any NPs in the sentence,
there are semantic cues to raising verbs even in the ambiguous sentence frame:
these cues come from the eventivity of the predicate of the lower clause and from
the animacy of the main clause subject. An inanimate subject paired with a stative
downstairs predicate yielded a raising verb 70% of the time.
3 Learning strategy
We have seen that there are cues to raising verbs in the input. The next question
is, do children attend to that information, and at what age do they do so? What
should the learning strategy look like?
There are three things that children should attend to in order to learn
raising verbs:
1. Children should attend to the semantic relationship between the upstairs subject
and the downstairs predicate. This is a good strategy because this relationship
is important for both raising and control structures. (In both raising and control
sentences, the main clause subject is the semantic subject of the downstairs
predicate.)
2. When children learn that inanimate things don’t have volition or intention, they
should attend to the semantic relationship between the upstairs subject and
upstairs predicate. This will allow children to distinguish raising from control
structures, since an inanimate subject is a poor subject of a control verb.
3. Children should use information from expletives.
Is there evidence that younger children attend to the semantic relation
between the upstairs subject and the downstairs predicate, and that older children
attend to semantic relations within the upper clause? There is some preliminary
evidence that this is so from an on-going experiment I am conducting. The task is
a version of the Truth-Value Judgment Task (Crain & McKee 1985). Children are
shown a picture, a puppet says something about the picture, and the child judges
whether the puppet’s comment was “good” (i.e. acceptable) or “silly”. If the child
judges the puppet’s utterance to be silly, the child is asked why it was silly.
Children were given 4 raising sentences, 4 control sentences and 8 interspersed
fillers. All test items (raising and control) contain an inanimate subject. Half of
the test sentences contain a predicate that is compatible with the (main clause)
subject; half of the test sentences contain a predicate that is incompatible with the
subject.
(20)
(21)
control
a. The door is trying to be purple.
b. The door is trying to be friendly.
raising
a. The hay seems to be on the ground.
b. The hay seems to be excited.
(compatible)
(incompatible)
(compatible)
(incompatible)
For our present purposes, the interesting items are those like (20a): a
control sentence with a compatible lower predicate. If the child attends only to the
relation between the subject and the downstairs predicate, the child should accept
the sentence (doors can be purple). If the child attends to the relation between the
subject and the upstairs predicate, the child should (like adults) reject the sentence
(doors can’t try to do or be anything). Indeed, with sentences of this type five
children with a mean age of 5;5 (years;months) attended to the relationship
between the subject and the upstairs predicate. Thus, when the downstairs
predicate was compatible with the subject (e.g. The door is trying to be purple),
children correctly rejected the sentence (judged it to be silly) 90% of the time.
(22)
Experimenter: #The door is trying to be purple
Child (5;5): “No, because you have to paint it”
Five younger children with a mean age of 3;3 showed a very different
pattern, accepting sentences like (20a) 100% of the time (judged to be good). Six
children with a mean age of 4;9 showed a mixed pattern, accepting sentences like
(20a) 50% of the time.
(23)
Experimenter: #The door is trying to be purple
Child (3;3—4;9): “Good”
Table 7: Children’s rate of attending to subject-predicate relations
mean age
door—try
door—be purple
unclear
5;5
90%
0%
10%
4;9
33.3%
50%
16.7%
3;3
0%
100%
0%
It appears that the incompatibility between inanimate subjects and control
verbs is acquired sometime around or after age 5. This (preliminary) result is
compatible with the idea that children first care about the relation between the
subject and the downstairs predicate. Please note that children in all age groups
had little or no trouble with raising sentences: children correctly accepted or
rejected raising sentences (e.g. The hay seems to be on the ground/#excited) at
least 70% of the time (age 3;3: 70% correct, age 4;9: 87.5% correct, age 5;5:
95.8% correct).
In other on-going work I am exploring children’s understanding of
expletive subjects. Interestingly, preliminary results suggest that children as
young as age 3;4 have no difficulty interpreting expletive it as an expletive (and
not as referential); some children, however, appear to overextend expletive it,
analyzing referential it as an expletive in certain cases. Further work is planned to
find out more about children’s interpretations of expletive and referential it.
Future work will involve teaching children a novel verb by using it in the
different sentential contexts I have used with adult English-speakers in the
experiments described above, and then testing for what children think the new
verb means, or seeing how they think the novel verb could be used.
4 Conclusions
The fill-in-the-blank experiments with adult English speakers demonstrated that
in spite of the hazards that raising verbs present for learners, there are some cues
in the sentence strings of English to suggest to a learner that the main verb of a
sentence is a raising verb. One of these kinds of cues, occurrence with an
expletive subject, requires the learner to reason across sentence frames: noticing
that a particular verb can occur with an expletive subject is useful for parsing the
ambiguous sentence frame only if the learner assumes that the verb occurring with
the expletive has the same syntactic properties as the (same) verb occurring with a
referential subject.
The other kind of cue had to do with semantic properties of the
constituents within the ambiguous frame. Since inanimate things do not have
intentions, and since most control verbs require an intentional subject, inanimate
subjects make poor subjects of control verbs.iii Thus, an inanimate subject may
serve as a cue to a raising structure even though the subject itself does not stand in
any semantic relationship with the raising verb.
In addition to subject animacy, eventivity of the downstairs predicate
plays a role in the likelihood of the main clause verb to be judged as a raising or a
control predicate. The reason for this is not entirely clear at present, but it appears
that (some) raising verbs are more compatible with a stative downstairs predicate
than an eventive one.
The main purpose of conducting the experiments with adults was to
determine whether there was information in the input that could lead a learner of
English to distinguish raising verbs from control verbs. The adult evidence
suggests that there are such cues, so the next step is to find out whether and at
what age children attend to those cues. Preliminary results of an experiment with
children ages 3-5;6 suggests that subject (in)animacy becomes a cue for
distinguishing raising from control structures around age 5: children below age 5
did not consistently reject control sentences in which the main clause subject was
inanimate and the lower predicate was a plausible predicate for that subject (and
3-year-olds consistently accepted such sentences). Further work on children’s
interpretations of expletives is planned. It is hoped that this and other
experimental work will shed light on how children acquire raising verbs.
Acknowledgements
This research was supported by a postdoctoral fellowship at the Institute for
Research in Cognitive Science at the University of Pennsylvania, funded by NSF
(NSF-STC 8920230). I am grateful to Lila Gleitman, Randy Hendrick, Jeff Lidz,
members of the Gleitman-Trueswell lab group at the University of Pennsylvania,
and audiences at the University of North Carolina, University of Delaware,
UCLA, the 26th Penn Linguistics Colloquium and CLS 38 for stimulating
discussion and very helpful suggestions. I claim responsibility for all errors and
shortcomings.
Notes
i
There were two lists of test items in each experiment (though each subject saw items from only
one list), and both lists were presented both forwards and backwards. No order or list effects were
found.
ii
There was a nonsignificant difference between the rates of raising verb responses to stative
predicates denoting physical states (be tall) and to those denoting mental states (know, love).
iii
Interestingly, inanimate subjects are likewise poor candidates for subjects of purpose
constructions, although adults often interpreted the ambiguous frame as a purpose construction.
References
Borer, Hagit and Kenneth Wexler. 1987. The maturation of syntax. Parameter setting, ed. by
Thomas Roeper and Edwin Williams. Dordrecht: Reidel.
Chomsky, Noam. 1981. Lectures on government and binding: The Pisa lectures. New York:
Mouton de Gruyter.
Crain, Stephen and Cecile McKee. 1985. The acquisition of structural restrictions on anaphora.
Proceedings of NELS 15, ed. by S. Berman, J. Choe, and J. McDonough. Amherst, MA: GLSA.
Fisher, Cynthia, Henry Gleitman and Lila R. Gleitman. 1991. On the semantic content of
subcategorization frames. Cognitive Psychology 23:331-392.
Fisher, Cynthia, D. Geoffrey Hall, Susan Rakowitz and Lila Gleitman. 1994. When it is better to
receive than to give: Syntactic and conceptual constraints on vocabulary growth. The acquisition
of the lexicon, ed. by Lila Gleitman and Barbara Landau. Cambridge, MA: MIT Press.
Frank, Robert. 1998. Structural complexity and the time course of grammatical development.
Cognition 66:249-301.
Gleitman, Lila. 1990. The structural sources of verb meanings. Language Acquisition 1: 3-55
Jackendoff, Ray. 1983. Semantics and cognition. Cambridge, MA: MIT Press.
Kako, Edward. 1998. The event semantics of syntactic structures. Ph.D. dissertation, University of
Pennsylvania.
Naigles, Letitia. 1996. The use of multiple frames in verb learning via syntactic bootstrapping.
Cognition 58: 221-251.
Naigles, Letitia, Henry Gleitman and Lila R. Gleitman. 1989. Syntactic bootstrapping in verb
acquisition: Evidence from comprehension. Language and cognition: A developmental
perspective, ed. by E. Drom. Norwood, NJ: Ablex.